1 /* Copyright (C) 2009 Red Hat, Inc.
2 * Copyright (C) 2006 Rusty Russell IBM Corporation
4 * Author: Michael S. Tsirkin <mst@redhat.com>
6 * Inspiration, some code, and most witty comments come from
7 * Documentation/virtual/lguest/lguest.c, by Rusty Russell
9 * This work is licensed under the terms of the GNU GPL, version 2.
11 * Generic code for virtio server in host kernel.
14 #include <linux/eventfd.h>
15 #include <linux/vhost.h>
16 #include <linux/uio.h>
18 #include <linux/mmu_context.h>
19 #include <linux/miscdevice.h>
20 #include <linux/mutex.h>
21 #include <linux/poll.h>
22 #include <linux/file.h>
23 #include <linux/highmem.h>
24 #include <linux/slab.h>
25 #include <linux/vmalloc.h>
26 #include <linux/kthread.h>
27 #include <linux/cgroup.h>
28 #include <linux/module.h>
29 #include <linux/sort.h>
30 #include <linux/sched/mm.h>
31 #include <linux/sched/signal.h>
32 #include <linux/interval_tree_generic.h>
33 #include <linux/nospec.h>
37 static ushort max_mem_regions = 64;
38 module_param(max_mem_regions, ushort, 0444);
39 MODULE_PARM_DESC(max_mem_regions,
40 "Maximum number of memory regions in memory map. (default: 64)");
41 static int max_iotlb_entries = 2048;
42 module_param(max_iotlb_entries, int, 0444);
43 MODULE_PARM_DESC(max_iotlb_entries,
44 "Maximum number of iotlb entries. (default: 2048)");
47 VHOST_MEMORY_F_LOG = 0x1,
50 #define vhost_used_event(vq) ((__virtio16 __user *)&vq->avail->ring[vq->num])
51 #define vhost_avail_event(vq) ((__virtio16 __user *)&vq->used->ring[vq->num])
53 INTERVAL_TREE_DEFINE(struct vhost_umem_node,
54 rb, __u64, __subtree_last,
55 START, LAST, static inline, vhost_umem_interval_tree);
57 #ifdef CONFIG_VHOST_CROSS_ENDIAN_LEGACY
58 static void vhost_disable_cross_endian(struct vhost_virtqueue *vq)
60 vq->user_be = !virtio_legacy_is_little_endian();
63 static void vhost_enable_cross_endian_big(struct vhost_virtqueue *vq)
68 static void vhost_enable_cross_endian_little(struct vhost_virtqueue *vq)
73 static long vhost_set_vring_endian(struct vhost_virtqueue *vq, int __user *argp)
75 struct vhost_vring_state s;
80 if (copy_from_user(&s, argp, sizeof(s)))
83 if (s.num != VHOST_VRING_LITTLE_ENDIAN &&
84 s.num != VHOST_VRING_BIG_ENDIAN)
87 if (s.num == VHOST_VRING_BIG_ENDIAN)
88 vhost_enable_cross_endian_big(vq);
90 vhost_enable_cross_endian_little(vq);
95 static long vhost_get_vring_endian(struct vhost_virtqueue *vq, u32 idx,
98 struct vhost_vring_state s = {
103 if (copy_to_user(argp, &s, sizeof(s)))
109 static void vhost_init_is_le(struct vhost_virtqueue *vq)
111 /* Note for legacy virtio: user_be is initialized at reset time
112 * according to the host endianness. If userspace does not set an
113 * explicit endianness, the default behavior is native endian, as
114 * expected by legacy virtio.
116 vq->is_le = vhost_has_feature(vq, VIRTIO_F_VERSION_1) || !vq->user_be;
119 static void vhost_disable_cross_endian(struct vhost_virtqueue *vq)
123 static long vhost_set_vring_endian(struct vhost_virtqueue *vq, int __user *argp)
128 static long vhost_get_vring_endian(struct vhost_virtqueue *vq, u32 idx,
134 static void vhost_init_is_le(struct vhost_virtqueue *vq)
136 vq->is_le = vhost_has_feature(vq, VIRTIO_F_VERSION_1)
137 || virtio_legacy_is_little_endian();
139 #endif /* CONFIG_VHOST_CROSS_ENDIAN_LEGACY */
141 static void vhost_reset_is_le(struct vhost_virtqueue *vq)
143 vhost_init_is_le(vq);
146 struct vhost_flush_struct {
147 struct vhost_work work;
148 struct completion wait_event;
151 static void vhost_flush_work(struct vhost_work *work)
153 struct vhost_flush_struct *s;
155 s = container_of(work, struct vhost_flush_struct, work);
156 complete(&s->wait_event);
159 static void vhost_poll_func(struct file *file, wait_queue_head_t *wqh,
162 struct vhost_poll *poll;
164 poll = container_of(pt, struct vhost_poll, table);
166 add_wait_queue(wqh, &poll->wait);
169 static int vhost_poll_wakeup(wait_queue_entry_t *wait, unsigned mode, int sync,
172 struct vhost_poll *poll = container_of(wait, struct vhost_poll, wait);
174 if (!(key_to_poll(key) & poll->mask))
177 vhost_poll_queue(poll);
181 void vhost_work_init(struct vhost_work *work, vhost_work_fn_t fn)
183 clear_bit(VHOST_WORK_QUEUED, &work->flags);
186 EXPORT_SYMBOL_GPL(vhost_work_init);
188 /* Init poll structure */
189 void vhost_poll_init(struct vhost_poll *poll, vhost_work_fn_t fn,
190 __poll_t mask, struct vhost_dev *dev)
192 init_waitqueue_func_entry(&poll->wait, vhost_poll_wakeup);
193 init_poll_funcptr(&poll->table, vhost_poll_func);
198 vhost_work_init(&poll->work, fn);
200 EXPORT_SYMBOL_GPL(vhost_poll_init);
202 /* Start polling a file. We add ourselves to file's wait queue. The caller must
203 * keep a reference to a file until after vhost_poll_stop is called. */
204 int vhost_poll_start(struct vhost_poll *poll, struct file *file)
212 mask = vfs_poll(file, &poll->table);
214 vhost_poll_wakeup(&poll->wait, 0, 0, poll_to_key(mask));
215 if (mask & EPOLLERR) {
216 vhost_poll_stop(poll);
222 EXPORT_SYMBOL_GPL(vhost_poll_start);
224 /* Stop polling a file. After this function returns, it becomes safe to drop the
225 * file reference. You must also flush afterwards. */
226 void vhost_poll_stop(struct vhost_poll *poll)
229 remove_wait_queue(poll->wqh, &poll->wait);
233 EXPORT_SYMBOL_GPL(vhost_poll_stop);
235 void vhost_work_flush(struct vhost_dev *dev, struct vhost_work *work)
237 struct vhost_flush_struct flush;
240 init_completion(&flush.wait_event);
241 vhost_work_init(&flush.work, vhost_flush_work);
243 vhost_work_queue(dev, &flush.work);
244 wait_for_completion(&flush.wait_event);
247 EXPORT_SYMBOL_GPL(vhost_work_flush);
249 /* Flush any work that has been scheduled. When calling this, don't hold any
250 * locks that are also used by the callback. */
251 void vhost_poll_flush(struct vhost_poll *poll)
253 vhost_work_flush(poll->dev, &poll->work);
255 EXPORT_SYMBOL_GPL(vhost_poll_flush);
257 void vhost_work_queue(struct vhost_dev *dev, struct vhost_work *work)
262 if (!test_and_set_bit(VHOST_WORK_QUEUED, &work->flags)) {
263 /* We can only add the work to the list after we're
264 * sure it was not in the list.
265 * test_and_set_bit() implies a memory barrier.
267 llist_add(&work->node, &dev->work_list);
268 wake_up_process(dev->worker);
271 EXPORT_SYMBOL_GPL(vhost_work_queue);
273 /* A lockless hint for busy polling code to exit the loop */
274 bool vhost_has_work(struct vhost_dev *dev)
276 return !llist_empty(&dev->work_list);
278 EXPORT_SYMBOL_GPL(vhost_has_work);
280 void vhost_poll_queue(struct vhost_poll *poll)
282 vhost_work_queue(poll->dev, &poll->work);
284 EXPORT_SYMBOL_GPL(vhost_poll_queue);
286 static void __vhost_vq_meta_reset(struct vhost_virtqueue *vq)
290 for (j = 0; j < VHOST_NUM_ADDRS; j++)
291 vq->meta_iotlb[j] = NULL;
294 static void vhost_vq_meta_reset(struct vhost_dev *d)
298 for (i = 0; i < d->nvqs; ++i)
299 __vhost_vq_meta_reset(d->vqs[i]);
302 static void vhost_vq_reset(struct vhost_dev *dev,
303 struct vhost_virtqueue *vq)
309 vq->last_avail_idx = 0;
311 vq->last_used_idx = 0;
312 vq->signalled_used = 0;
313 vq->signalled_used_valid = false;
315 vq->log_used = false;
316 vq->log_addr = -1ull;
317 vq->private_data = NULL;
318 vq->acked_features = 0;
319 vq->acked_backend_features = 0;
321 vq->error_ctx = NULL;
325 vhost_reset_is_le(vq);
326 vhost_disable_cross_endian(vq);
327 vq->busyloop_timeout = 0;
330 __vhost_vq_meta_reset(vq);
333 static int vhost_worker(void *data)
335 struct vhost_dev *dev = data;
336 struct vhost_work *work, *work_next;
337 struct llist_node *node;
338 mm_segment_t oldfs = get_fs();
344 /* mb paired w/ kthread_stop */
345 set_current_state(TASK_INTERRUPTIBLE);
347 if (kthread_should_stop()) {
348 __set_current_state(TASK_RUNNING);
352 node = llist_del_all(&dev->work_list);
356 node = llist_reverse_order(node);
357 /* make sure flag is seen after deletion */
359 llist_for_each_entry_safe(work, work_next, node, node) {
360 clear_bit(VHOST_WORK_QUEUED, &work->flags);
361 __set_current_state(TASK_RUNNING);
372 static void vhost_vq_free_iovecs(struct vhost_virtqueue *vq)
382 /* Helper to allocate iovec buffers for all vqs. */
383 static long vhost_dev_alloc_iovecs(struct vhost_dev *dev)
385 struct vhost_virtqueue *vq;
388 for (i = 0; i < dev->nvqs; ++i) {
390 vq->indirect = kmalloc_array(UIO_MAXIOV,
391 sizeof(*vq->indirect),
393 vq->log = kmalloc_array(UIO_MAXIOV, sizeof(*vq->log),
395 vq->heads = kmalloc_array(UIO_MAXIOV, sizeof(*vq->heads),
397 if (!vq->indirect || !vq->log || !vq->heads)
404 vhost_vq_free_iovecs(dev->vqs[i]);
408 static void vhost_dev_free_iovecs(struct vhost_dev *dev)
412 for (i = 0; i < dev->nvqs; ++i)
413 vhost_vq_free_iovecs(dev->vqs[i]);
416 void vhost_dev_init(struct vhost_dev *dev,
417 struct vhost_virtqueue **vqs, int nvqs)
419 struct vhost_virtqueue *vq;
424 mutex_init(&dev->mutex);
430 init_llist_head(&dev->work_list);
431 init_waitqueue_head(&dev->wait);
432 INIT_LIST_HEAD(&dev->read_list);
433 INIT_LIST_HEAD(&dev->pending_list);
434 spin_lock_init(&dev->iotlb_lock);
437 for (i = 0; i < dev->nvqs; ++i) {
443 mutex_init(&vq->mutex);
444 vhost_vq_reset(dev, vq);
446 vhost_poll_init(&vq->poll, vq->handle_kick,
450 EXPORT_SYMBOL_GPL(vhost_dev_init);
452 /* Caller should have device mutex */
453 long vhost_dev_check_owner(struct vhost_dev *dev)
455 /* Are you the owner? If not, I don't think you mean to do that */
456 return dev->mm == current->mm ? 0 : -EPERM;
458 EXPORT_SYMBOL_GPL(vhost_dev_check_owner);
460 struct vhost_attach_cgroups_struct {
461 struct vhost_work work;
462 struct task_struct *owner;
466 static void vhost_attach_cgroups_work(struct vhost_work *work)
468 struct vhost_attach_cgroups_struct *s;
470 s = container_of(work, struct vhost_attach_cgroups_struct, work);
471 s->ret = cgroup_attach_task_all(s->owner, current);
474 static int vhost_attach_cgroups(struct vhost_dev *dev)
476 struct vhost_attach_cgroups_struct attach;
478 attach.owner = current;
479 vhost_work_init(&attach.work, vhost_attach_cgroups_work);
480 vhost_work_queue(dev, &attach.work);
481 vhost_work_flush(dev, &attach.work);
485 /* Caller should have device mutex */
486 bool vhost_dev_has_owner(struct vhost_dev *dev)
490 EXPORT_SYMBOL_GPL(vhost_dev_has_owner);
492 /* Caller should have device mutex */
493 long vhost_dev_set_owner(struct vhost_dev *dev)
495 struct task_struct *worker;
498 /* Is there an owner already? */
499 if (vhost_dev_has_owner(dev)) {
504 /* No owner, become one */
505 dev->mm = get_task_mm(current);
506 worker = kthread_create(vhost_worker, dev, "vhost-%d", current->pid);
507 if (IS_ERR(worker)) {
508 err = PTR_ERR(worker);
512 dev->worker = worker;
513 wake_up_process(worker); /* avoid contributing to loadavg */
515 err = vhost_attach_cgroups(dev);
519 err = vhost_dev_alloc_iovecs(dev);
525 kthread_stop(worker);
534 EXPORT_SYMBOL_GPL(vhost_dev_set_owner);
536 struct vhost_umem *vhost_dev_reset_owner_prepare(void)
538 return kvzalloc(sizeof(struct vhost_umem), GFP_KERNEL);
540 EXPORT_SYMBOL_GPL(vhost_dev_reset_owner_prepare);
542 /* Caller should have device mutex */
543 void vhost_dev_reset_owner(struct vhost_dev *dev, struct vhost_umem *umem)
547 vhost_dev_cleanup(dev);
549 /* Restore memory to default empty mapping. */
550 INIT_LIST_HEAD(&umem->umem_list);
552 /* We don't need VQ locks below since vhost_dev_cleanup makes sure
553 * VQs aren't running.
555 for (i = 0; i < dev->nvqs; ++i)
556 dev->vqs[i]->umem = umem;
558 EXPORT_SYMBOL_GPL(vhost_dev_reset_owner);
560 void vhost_dev_stop(struct vhost_dev *dev)
564 for (i = 0; i < dev->nvqs; ++i) {
565 if (dev->vqs[i]->kick && dev->vqs[i]->handle_kick) {
566 vhost_poll_stop(&dev->vqs[i]->poll);
567 vhost_poll_flush(&dev->vqs[i]->poll);
571 EXPORT_SYMBOL_GPL(vhost_dev_stop);
573 static void vhost_umem_free(struct vhost_umem *umem,
574 struct vhost_umem_node *node)
576 vhost_umem_interval_tree_remove(node, &umem->umem_tree);
577 list_del(&node->link);
582 static void vhost_umem_clean(struct vhost_umem *umem)
584 struct vhost_umem_node *node, *tmp;
589 list_for_each_entry_safe(node, tmp, &umem->umem_list, link)
590 vhost_umem_free(umem, node);
595 static void vhost_clear_msg(struct vhost_dev *dev)
597 struct vhost_msg_node *node, *n;
599 spin_lock(&dev->iotlb_lock);
601 list_for_each_entry_safe(node, n, &dev->read_list, node) {
602 list_del(&node->node);
606 list_for_each_entry_safe(node, n, &dev->pending_list, node) {
607 list_del(&node->node);
611 spin_unlock(&dev->iotlb_lock);
614 void vhost_dev_cleanup(struct vhost_dev *dev)
618 for (i = 0; i < dev->nvqs; ++i) {
619 if (dev->vqs[i]->error_ctx)
620 eventfd_ctx_put(dev->vqs[i]->error_ctx);
621 if (dev->vqs[i]->kick)
622 fput(dev->vqs[i]->kick);
623 if (dev->vqs[i]->call_ctx)
624 eventfd_ctx_put(dev->vqs[i]->call_ctx);
625 vhost_vq_reset(dev, dev->vqs[i]);
627 vhost_dev_free_iovecs(dev);
629 eventfd_ctx_put(dev->log_ctx);
631 /* No one will access memory at this point */
632 vhost_umem_clean(dev->umem);
634 vhost_umem_clean(dev->iotlb);
636 vhost_clear_msg(dev);
637 wake_up_interruptible_poll(&dev->wait, EPOLLIN | EPOLLRDNORM);
638 WARN_ON(!llist_empty(&dev->work_list));
640 kthread_stop(dev->worker);
647 EXPORT_SYMBOL_GPL(vhost_dev_cleanup);
649 static bool log_access_ok(void __user *log_base, u64 addr, unsigned long sz)
651 u64 a = addr / VHOST_PAGE_SIZE / 8;
653 /* Make sure 64 bit math will not overflow. */
654 if (a > ULONG_MAX - (unsigned long)log_base ||
655 a + (unsigned long)log_base > ULONG_MAX)
658 return access_ok(VERIFY_WRITE, log_base + a,
659 (sz + VHOST_PAGE_SIZE * 8 - 1) / VHOST_PAGE_SIZE / 8);
662 static bool vhost_overflow(u64 uaddr, u64 size)
664 /* Make sure 64 bit math will not overflow. */
665 return uaddr > ULONG_MAX || size > ULONG_MAX || uaddr > ULONG_MAX - size;
668 /* Caller should have vq mutex and device mutex. */
669 static bool vq_memory_access_ok(void __user *log_base, struct vhost_umem *umem,
672 struct vhost_umem_node *node;
677 list_for_each_entry(node, &umem->umem_list, link) {
678 unsigned long a = node->userspace_addr;
680 if (vhost_overflow(node->userspace_addr, node->size))
684 if (!access_ok(VERIFY_WRITE, (void __user *)a,
687 else if (log_all && !log_access_ok(log_base,
695 static inline void __user *vhost_vq_meta_fetch(struct vhost_virtqueue *vq,
696 u64 addr, unsigned int size,
699 const struct vhost_umem_node *node = vq->meta_iotlb[type];
704 return (void *)(uintptr_t)(node->userspace_addr + addr - node->start);
707 /* Can we switch to this memory table? */
708 /* Caller should have device mutex but not vq mutex */
709 static bool memory_access_ok(struct vhost_dev *d, struct vhost_umem *umem,
714 for (i = 0; i < d->nvqs; ++i) {
718 mutex_lock(&d->vqs[i]->mutex);
719 log = log_all || vhost_has_feature(d->vqs[i], VHOST_F_LOG_ALL);
720 /* If ring is inactive, will check when it's enabled. */
721 if (d->vqs[i]->private_data)
722 ok = vq_memory_access_ok(d->vqs[i]->log_base,
726 mutex_unlock(&d->vqs[i]->mutex);
733 static int translate_desc(struct vhost_virtqueue *vq, u64 addr, u32 len,
734 struct iovec iov[], int iov_size, int access);
736 static int vhost_copy_to_user(struct vhost_virtqueue *vq, void __user *to,
737 const void *from, unsigned size)
742 return __copy_to_user(to, from, size);
744 /* This function should be called after iotlb
745 * prefetch, which means we're sure that all vq
746 * could be access through iotlb. So -EAGAIN should
747 * not happen in this case.
750 void __user *uaddr = vhost_vq_meta_fetch(vq,
751 (u64)(uintptr_t)to, size,
755 return __copy_to_user(uaddr, from, size);
757 ret = translate_desc(vq, (u64)(uintptr_t)to, size, vq->iotlb_iov,
758 ARRAY_SIZE(vq->iotlb_iov),
762 iov_iter_init(&t, WRITE, vq->iotlb_iov, ret, size);
763 ret = copy_to_iter(from, size, &t);
771 static int vhost_copy_from_user(struct vhost_virtqueue *vq, void *to,
772 void __user *from, unsigned size)
777 return __copy_from_user(to, from, size);
779 /* This function should be called after iotlb
780 * prefetch, which means we're sure that vq
781 * could be access through iotlb. So -EAGAIN should
782 * not happen in this case.
784 void __user *uaddr = vhost_vq_meta_fetch(vq,
785 (u64)(uintptr_t)from, size,
790 return __copy_from_user(to, uaddr, size);
792 ret = translate_desc(vq, (u64)(uintptr_t)from, size, vq->iotlb_iov,
793 ARRAY_SIZE(vq->iotlb_iov),
796 vq_err(vq, "IOTLB translation failure: uaddr "
797 "%p size 0x%llx\n", from,
798 (unsigned long long) size);
801 iov_iter_init(&f, READ, vq->iotlb_iov, ret, size);
802 ret = copy_from_iter(to, size, &f);
811 static void __user *__vhost_get_user_slow(struct vhost_virtqueue *vq,
812 void __user *addr, unsigned int size,
817 ret = translate_desc(vq, (u64)(uintptr_t)addr, size, vq->iotlb_iov,
818 ARRAY_SIZE(vq->iotlb_iov),
821 vq_err(vq, "IOTLB translation failure: uaddr "
822 "%p size 0x%llx\n", addr,
823 (unsigned long long) size);
827 if (ret != 1 || vq->iotlb_iov[0].iov_len != size) {
828 vq_err(vq, "Non atomic userspace memory access: uaddr "
829 "%p size 0x%llx\n", addr,
830 (unsigned long long) size);
834 return vq->iotlb_iov[0].iov_base;
837 /* This function should be called after iotlb
838 * prefetch, which means we're sure that vq
839 * could be access through iotlb. So -EAGAIN should
840 * not happen in this case.
842 static inline void __user *__vhost_get_user(struct vhost_virtqueue *vq,
843 void *addr, unsigned int size,
846 void __user *uaddr = vhost_vq_meta_fetch(vq,
847 (u64)(uintptr_t)addr, size, type);
851 return __vhost_get_user_slow(vq, addr, size, type);
854 #define vhost_put_user(vq, x, ptr) \
858 ret = __put_user(x, ptr); \
860 __typeof__(ptr) to = \
861 (__typeof__(ptr)) __vhost_get_user(vq, ptr, \
862 sizeof(*ptr), VHOST_ADDR_USED); \
864 ret = __put_user(x, to); \
871 #define vhost_get_user(vq, x, ptr, type) \
875 ret = __get_user(x, ptr); \
877 __typeof__(ptr) from = \
878 (__typeof__(ptr)) __vhost_get_user(vq, ptr, \
882 ret = __get_user(x, from); \
889 #define vhost_get_avail(vq, x, ptr) \
890 vhost_get_user(vq, x, ptr, VHOST_ADDR_AVAIL)
892 #define vhost_get_used(vq, x, ptr) \
893 vhost_get_user(vq, x, ptr, VHOST_ADDR_USED)
895 static void vhost_dev_lock_vqs(struct vhost_dev *d)
898 for (i = 0; i < d->nvqs; ++i)
899 mutex_lock_nested(&d->vqs[i]->mutex, i);
902 static void vhost_dev_unlock_vqs(struct vhost_dev *d)
905 for (i = 0; i < d->nvqs; ++i)
906 mutex_unlock(&d->vqs[i]->mutex);
909 static int vhost_new_umem_range(struct vhost_umem *umem,
910 u64 start, u64 size, u64 end,
911 u64 userspace_addr, int perm)
913 struct vhost_umem_node *tmp, *node = kmalloc(sizeof(*node), GFP_ATOMIC);
918 if (umem->numem == max_iotlb_entries) {
919 tmp = list_first_entry(&umem->umem_list, typeof(*tmp), link);
920 vhost_umem_free(umem, tmp);
926 node->userspace_addr = userspace_addr;
928 INIT_LIST_HEAD(&node->link);
929 list_add_tail(&node->link, &umem->umem_list);
930 vhost_umem_interval_tree_insert(node, &umem->umem_tree);
936 static void vhost_del_umem_range(struct vhost_umem *umem,
939 struct vhost_umem_node *node;
941 while ((node = vhost_umem_interval_tree_iter_first(&umem->umem_tree,
943 vhost_umem_free(umem, node);
946 static void vhost_iotlb_notify_vq(struct vhost_dev *d,
947 struct vhost_iotlb_msg *msg)
949 struct vhost_msg_node *node, *n;
951 spin_lock(&d->iotlb_lock);
953 list_for_each_entry_safe(node, n, &d->pending_list, node) {
954 struct vhost_iotlb_msg *vq_msg = &node->msg.iotlb;
955 if (msg->iova <= vq_msg->iova &&
956 msg->iova + msg->size - 1 >= vq_msg->iova &&
957 vq_msg->type == VHOST_IOTLB_MISS) {
958 vhost_poll_queue(&node->vq->poll);
959 list_del(&node->node);
964 spin_unlock(&d->iotlb_lock);
967 static bool umem_access_ok(u64 uaddr, u64 size, int access)
969 unsigned long a = uaddr;
971 /* Make sure 64 bit math will not overflow. */
972 if (vhost_overflow(uaddr, size))
975 if ((access & VHOST_ACCESS_RO) &&
976 !access_ok(VERIFY_READ, (void __user *)a, size))
978 if ((access & VHOST_ACCESS_WO) &&
979 !access_ok(VERIFY_WRITE, (void __user *)a, size))
984 static int vhost_process_iotlb_msg(struct vhost_dev *dev,
985 struct vhost_iotlb_msg *msg)
989 mutex_lock(&dev->mutex);
990 vhost_dev_lock_vqs(dev);
992 case VHOST_IOTLB_UPDATE:
997 if (!umem_access_ok(msg->uaddr, msg->size, msg->perm)) {
1001 vhost_vq_meta_reset(dev);
1002 if (vhost_new_umem_range(dev->iotlb, msg->iova, msg->size,
1003 msg->iova + msg->size - 1,
1004 msg->uaddr, msg->perm)) {
1008 vhost_iotlb_notify_vq(dev, msg);
1010 case VHOST_IOTLB_INVALIDATE:
1015 vhost_vq_meta_reset(dev);
1016 vhost_del_umem_range(dev->iotlb, msg->iova,
1017 msg->iova + msg->size - 1);
1024 vhost_dev_unlock_vqs(dev);
1025 mutex_unlock(&dev->mutex);
1029 ssize_t vhost_chr_write_iter(struct vhost_dev *dev,
1030 struct iov_iter *from)
1032 struct vhost_iotlb_msg msg;
1036 ret = copy_from_iter(&type, sizeof(type), from);
1037 if (ret != sizeof(type))
1041 case VHOST_IOTLB_MSG:
1042 /* There maybe a hole after type for V1 message type,
1045 offset = offsetof(struct vhost_msg, iotlb) - sizeof(int);
1047 case VHOST_IOTLB_MSG_V2:
1048 offset = sizeof(__u32);
1055 iov_iter_advance(from, offset);
1056 ret = copy_from_iter(&msg, sizeof(msg), from);
1057 if (ret != sizeof(msg))
1059 if (vhost_process_iotlb_msg(dev, &msg)) {
1064 ret = (type == VHOST_IOTLB_MSG) ? sizeof(struct vhost_msg) :
1065 sizeof(struct vhost_msg_v2);
1069 EXPORT_SYMBOL(vhost_chr_write_iter);
1071 __poll_t vhost_chr_poll(struct file *file, struct vhost_dev *dev,
1076 poll_wait(file, &dev->wait, wait);
1078 if (!list_empty(&dev->read_list))
1079 mask |= EPOLLIN | EPOLLRDNORM;
1083 EXPORT_SYMBOL(vhost_chr_poll);
1085 ssize_t vhost_chr_read_iter(struct vhost_dev *dev, struct iov_iter *to,
1089 struct vhost_msg_node *node;
1091 unsigned size = sizeof(struct vhost_msg);
1093 if (iov_iter_count(to) < size)
1098 prepare_to_wait(&dev->wait, &wait,
1099 TASK_INTERRUPTIBLE);
1101 node = vhost_dequeue_msg(dev, &dev->read_list);
1108 if (signal_pending(current)) {
1121 finish_wait(&dev->wait, &wait);
1124 struct vhost_iotlb_msg *msg;
1125 void *start = &node->msg;
1127 switch (node->msg.type) {
1128 case VHOST_IOTLB_MSG:
1129 size = sizeof(node->msg);
1130 msg = &node->msg.iotlb;
1132 case VHOST_IOTLB_MSG_V2:
1133 size = sizeof(node->msg_v2);
1134 msg = &node->msg_v2.iotlb;
1141 ret = copy_to_iter(start, size, to);
1142 if (ret != size || msg->type != VHOST_IOTLB_MISS) {
1146 vhost_enqueue_msg(dev, &dev->pending_list, node);
1151 EXPORT_SYMBOL_GPL(vhost_chr_read_iter);
1153 static int vhost_iotlb_miss(struct vhost_virtqueue *vq, u64 iova, int access)
1155 struct vhost_dev *dev = vq->dev;
1156 struct vhost_msg_node *node;
1157 struct vhost_iotlb_msg *msg;
1158 bool v2 = vhost_backend_has_feature(vq, VHOST_BACKEND_F_IOTLB_MSG_V2);
1160 node = vhost_new_msg(vq, v2 ? VHOST_IOTLB_MSG_V2 : VHOST_IOTLB_MSG);
1165 node->msg_v2.type = VHOST_IOTLB_MSG_V2;
1166 msg = &node->msg_v2.iotlb;
1168 msg = &node->msg.iotlb;
1171 msg->type = VHOST_IOTLB_MISS;
1175 vhost_enqueue_msg(dev, &dev->read_list, node);
1180 static bool vq_access_ok(struct vhost_virtqueue *vq, unsigned int num,
1181 struct vring_desc __user *desc,
1182 struct vring_avail __user *avail,
1183 struct vring_used __user *used)
1186 size_t s = vhost_has_feature(vq, VIRTIO_RING_F_EVENT_IDX) ? 2 : 0;
1188 return access_ok(VERIFY_READ, desc, num * sizeof *desc) &&
1189 access_ok(VERIFY_READ, avail,
1190 sizeof *avail + num * sizeof *avail->ring + s) &&
1191 access_ok(VERIFY_WRITE, used,
1192 sizeof *used + num * sizeof *used->ring + s);
1195 static void vhost_vq_meta_update(struct vhost_virtqueue *vq,
1196 const struct vhost_umem_node *node,
1199 int access = (type == VHOST_ADDR_USED) ?
1200 VHOST_ACCESS_WO : VHOST_ACCESS_RO;
1202 if (likely(node->perm & access))
1203 vq->meta_iotlb[type] = node;
1206 static bool iotlb_access_ok(struct vhost_virtqueue *vq,
1207 int access, u64 addr, u64 len, int type)
1209 const struct vhost_umem_node *node;
1210 struct vhost_umem *umem = vq->iotlb;
1211 u64 s = 0, size, orig_addr = addr, last = addr + len - 1;
1213 if (vhost_vq_meta_fetch(vq, addr, len, type))
1217 node = vhost_umem_interval_tree_iter_first(&umem->umem_tree,
1220 if (node == NULL || node->start > addr) {
1221 vhost_iotlb_miss(vq, addr, access);
1223 } else if (!(node->perm & access)) {
1224 /* Report the possible access violation by
1225 * request another translation from userspace.
1230 size = node->size - addr + node->start;
1232 if (orig_addr == addr && size >= len)
1233 vhost_vq_meta_update(vq, node, type);
1242 int vq_iotlb_prefetch(struct vhost_virtqueue *vq)
1244 size_t s = vhost_has_feature(vq, VIRTIO_RING_F_EVENT_IDX) ? 2 : 0;
1245 unsigned int num = vq->num;
1250 return iotlb_access_ok(vq, VHOST_ACCESS_RO, (u64)(uintptr_t)vq->desc,
1251 num * sizeof(*vq->desc), VHOST_ADDR_DESC) &&
1252 iotlb_access_ok(vq, VHOST_ACCESS_RO, (u64)(uintptr_t)vq->avail,
1254 num * sizeof(*vq->avail->ring) + s,
1255 VHOST_ADDR_AVAIL) &&
1256 iotlb_access_ok(vq, VHOST_ACCESS_WO, (u64)(uintptr_t)vq->used,
1258 num * sizeof(*vq->used->ring) + s,
1261 EXPORT_SYMBOL_GPL(vq_iotlb_prefetch);
1263 /* Can we log writes? */
1264 /* Caller should have device mutex but not vq mutex */
1265 bool vhost_log_access_ok(struct vhost_dev *dev)
1267 return memory_access_ok(dev, dev->umem, 1);
1269 EXPORT_SYMBOL_GPL(vhost_log_access_ok);
1271 /* Verify access for write logging. */
1272 /* Caller should have vq mutex and device mutex */
1273 static bool vq_log_access_ok(struct vhost_virtqueue *vq,
1274 void __user *log_base)
1276 size_t s = vhost_has_feature(vq, VIRTIO_RING_F_EVENT_IDX) ? 2 : 0;
1278 return vq_memory_access_ok(log_base, vq->umem,
1279 vhost_has_feature(vq, VHOST_F_LOG_ALL)) &&
1280 (!vq->log_used || log_access_ok(log_base, vq->log_addr,
1282 vq->num * sizeof *vq->used->ring + s));
1285 /* Can we start vq? */
1286 /* Caller should have vq mutex and device mutex */
1287 bool vhost_vq_access_ok(struct vhost_virtqueue *vq)
1289 if (!vq_log_access_ok(vq, vq->log_base))
1292 /* Access validation occurs at prefetch time with IOTLB */
1296 return vq_access_ok(vq, vq->num, vq->desc, vq->avail, vq->used);
1298 EXPORT_SYMBOL_GPL(vhost_vq_access_ok);
1300 static struct vhost_umem *vhost_umem_alloc(void)
1302 struct vhost_umem *umem = kvzalloc(sizeof(*umem), GFP_KERNEL);
1307 umem->umem_tree = RB_ROOT_CACHED;
1309 INIT_LIST_HEAD(&umem->umem_list);
1314 static long vhost_set_memory(struct vhost_dev *d, struct vhost_memory __user *m)
1316 struct vhost_memory mem, *newmem;
1317 struct vhost_memory_region *region;
1318 struct vhost_umem *newumem, *oldumem;
1319 unsigned long size = offsetof(struct vhost_memory, regions);
1322 if (copy_from_user(&mem, m, size))
1326 if (mem.nregions > max_mem_regions)
1328 newmem = kvzalloc(struct_size(newmem, regions, mem.nregions),
1333 memcpy(newmem, &mem, size);
1334 if (copy_from_user(newmem->regions, m->regions,
1335 mem.nregions * sizeof *m->regions)) {
1340 newumem = vhost_umem_alloc();
1346 for (region = newmem->regions;
1347 region < newmem->regions + mem.nregions;
1349 if (vhost_new_umem_range(newumem,
1350 region->guest_phys_addr,
1351 region->memory_size,
1352 region->guest_phys_addr +
1353 region->memory_size - 1,
1354 region->userspace_addr,
1359 if (!memory_access_ok(d, newumem, 0))
1365 /* All memory accesses are done under some VQ mutex. */
1366 for (i = 0; i < d->nvqs; ++i) {
1367 mutex_lock(&d->vqs[i]->mutex);
1368 d->vqs[i]->umem = newumem;
1369 mutex_unlock(&d->vqs[i]->mutex);
1373 vhost_umem_clean(oldumem);
1377 vhost_umem_clean(newumem);
1382 long vhost_vring_ioctl(struct vhost_dev *d, unsigned int ioctl, void __user *argp)
1384 struct file *eventfp, *filep = NULL;
1385 bool pollstart = false, pollstop = false;
1386 struct eventfd_ctx *ctx = NULL;
1387 u32 __user *idxp = argp;
1388 struct vhost_virtqueue *vq;
1389 struct vhost_vring_state s;
1390 struct vhost_vring_file f;
1391 struct vhost_vring_addr a;
1395 r = get_user(idx, idxp);
1401 idx = array_index_nospec(idx, d->nvqs);
1404 mutex_lock(&vq->mutex);
1407 case VHOST_SET_VRING_NUM:
1408 /* Resizing ring with an active backend?
1409 * You don't want to do that. */
1410 if (vq->private_data) {
1414 if (copy_from_user(&s, argp, sizeof s)) {
1418 if (!s.num || s.num > 0xffff || (s.num & (s.num - 1))) {
1424 case VHOST_SET_VRING_BASE:
1425 /* Moving base with an active backend?
1426 * You don't want to do that. */
1427 if (vq->private_data) {
1431 if (copy_from_user(&s, argp, sizeof s)) {
1435 if (s.num > 0xffff) {
1439 vq->last_avail_idx = s.num;
1440 /* Forget the cached index value. */
1441 vq->avail_idx = vq->last_avail_idx;
1443 case VHOST_GET_VRING_BASE:
1445 s.num = vq->last_avail_idx;
1446 if (copy_to_user(argp, &s, sizeof s))
1449 case VHOST_SET_VRING_ADDR:
1450 if (copy_from_user(&a, argp, sizeof a)) {
1454 if (a.flags & ~(0x1 << VHOST_VRING_F_LOG)) {
1458 /* For 32bit, verify that the top 32bits of the user
1459 data are set to zero. */
1460 if ((u64)(unsigned long)a.desc_user_addr != a.desc_user_addr ||
1461 (u64)(unsigned long)a.used_user_addr != a.used_user_addr ||
1462 (u64)(unsigned long)a.avail_user_addr != a.avail_user_addr) {
1467 /* Make sure it's safe to cast pointers to vring types. */
1468 BUILD_BUG_ON(__alignof__ *vq->avail > VRING_AVAIL_ALIGN_SIZE);
1469 BUILD_BUG_ON(__alignof__ *vq->used > VRING_USED_ALIGN_SIZE);
1470 if ((a.avail_user_addr & (VRING_AVAIL_ALIGN_SIZE - 1)) ||
1471 (a.used_user_addr & (VRING_USED_ALIGN_SIZE - 1)) ||
1472 (a.log_guest_addr & (VRING_USED_ALIGN_SIZE - 1))) {
1477 /* We only verify access here if backend is configured.
1478 * If it is not, we don't as size might not have been setup.
1479 * We will verify when backend is configured. */
1480 if (vq->private_data) {
1481 if (!vq_access_ok(vq, vq->num,
1482 (void __user *)(unsigned long)a.desc_user_addr,
1483 (void __user *)(unsigned long)a.avail_user_addr,
1484 (void __user *)(unsigned long)a.used_user_addr)) {
1489 /* Also validate log access for used ring if enabled. */
1490 if ((a.flags & (0x1 << VHOST_VRING_F_LOG)) &&
1491 !log_access_ok(vq->log_base, a.log_guest_addr,
1493 vq->num * sizeof *vq->used->ring)) {
1499 vq->log_used = !!(a.flags & (0x1 << VHOST_VRING_F_LOG));
1500 vq->desc = (void __user *)(unsigned long)a.desc_user_addr;
1501 vq->avail = (void __user *)(unsigned long)a.avail_user_addr;
1502 vq->log_addr = a.log_guest_addr;
1503 vq->used = (void __user *)(unsigned long)a.used_user_addr;
1505 case VHOST_SET_VRING_KICK:
1506 if (copy_from_user(&f, argp, sizeof f)) {
1510 eventfp = f.fd == -1 ? NULL : eventfd_fget(f.fd);
1511 if (IS_ERR(eventfp)) {
1512 r = PTR_ERR(eventfp);
1515 if (eventfp != vq->kick) {
1516 pollstop = (filep = vq->kick) != NULL;
1517 pollstart = (vq->kick = eventfp) != NULL;
1521 case VHOST_SET_VRING_CALL:
1522 if (copy_from_user(&f, argp, sizeof f)) {
1526 ctx = f.fd == -1 ? NULL : eventfd_ctx_fdget(f.fd);
1531 swap(ctx, vq->call_ctx);
1533 case VHOST_SET_VRING_ERR:
1534 if (copy_from_user(&f, argp, sizeof f)) {
1538 ctx = f.fd == -1 ? NULL : eventfd_ctx_fdget(f.fd);
1543 swap(ctx, vq->error_ctx);
1545 case VHOST_SET_VRING_ENDIAN:
1546 r = vhost_set_vring_endian(vq, argp);
1548 case VHOST_GET_VRING_ENDIAN:
1549 r = vhost_get_vring_endian(vq, idx, argp);
1551 case VHOST_SET_VRING_BUSYLOOP_TIMEOUT:
1552 if (copy_from_user(&s, argp, sizeof(s))) {
1556 vq->busyloop_timeout = s.num;
1558 case VHOST_GET_VRING_BUSYLOOP_TIMEOUT:
1560 s.num = vq->busyloop_timeout;
1561 if (copy_to_user(argp, &s, sizeof(s)))
1568 if (pollstop && vq->handle_kick)
1569 vhost_poll_stop(&vq->poll);
1571 if (!IS_ERR_OR_NULL(ctx))
1572 eventfd_ctx_put(ctx);
1576 if (pollstart && vq->handle_kick)
1577 r = vhost_poll_start(&vq->poll, vq->kick);
1579 mutex_unlock(&vq->mutex);
1581 if (pollstop && vq->handle_kick)
1582 vhost_poll_flush(&vq->poll);
1585 EXPORT_SYMBOL_GPL(vhost_vring_ioctl);
1587 int vhost_init_device_iotlb(struct vhost_dev *d, bool enabled)
1589 struct vhost_umem *niotlb, *oiotlb;
1592 niotlb = vhost_umem_alloc();
1599 for (i = 0; i < d->nvqs; ++i) {
1600 struct vhost_virtqueue *vq = d->vqs[i];
1602 mutex_lock(&vq->mutex);
1604 __vhost_vq_meta_reset(vq);
1605 mutex_unlock(&vq->mutex);
1608 vhost_umem_clean(oiotlb);
1612 EXPORT_SYMBOL_GPL(vhost_init_device_iotlb);
1614 /* Caller must have device mutex */
1615 long vhost_dev_ioctl(struct vhost_dev *d, unsigned int ioctl, void __user *argp)
1617 struct eventfd_ctx *ctx;
1622 /* If you are not the owner, you can become one */
1623 if (ioctl == VHOST_SET_OWNER) {
1624 r = vhost_dev_set_owner(d);
1628 /* You must be the owner to do anything else */
1629 r = vhost_dev_check_owner(d);
1634 case VHOST_SET_MEM_TABLE:
1635 r = vhost_set_memory(d, argp);
1637 case VHOST_SET_LOG_BASE:
1638 if (copy_from_user(&p, argp, sizeof p)) {
1642 if ((u64)(unsigned long)p != p) {
1646 for (i = 0; i < d->nvqs; ++i) {
1647 struct vhost_virtqueue *vq;
1648 void __user *base = (void __user *)(unsigned long)p;
1650 mutex_lock(&vq->mutex);
1651 /* If ring is inactive, will check when it's enabled. */
1652 if (vq->private_data && !vq_log_access_ok(vq, base))
1655 vq->log_base = base;
1656 mutex_unlock(&vq->mutex);
1659 case VHOST_SET_LOG_FD:
1660 r = get_user(fd, (int __user *)argp);
1663 ctx = fd == -1 ? NULL : eventfd_ctx_fdget(fd);
1668 swap(ctx, d->log_ctx);
1669 for (i = 0; i < d->nvqs; ++i) {
1670 mutex_lock(&d->vqs[i]->mutex);
1671 d->vqs[i]->log_ctx = d->log_ctx;
1672 mutex_unlock(&d->vqs[i]->mutex);
1675 eventfd_ctx_put(ctx);
1684 EXPORT_SYMBOL_GPL(vhost_dev_ioctl);
1686 /* TODO: This is really inefficient. We need something like get_user()
1687 * (instruction directly accesses the data, with an exception table entry
1688 * returning -EFAULT). See Documentation/x86/exception-tables.txt.
1690 static int set_bit_to_user(int nr, void __user *addr)
1692 unsigned long log = (unsigned long)addr;
1695 int bit = nr + (log % PAGE_SIZE) * 8;
1698 r = get_user_pages_fast(log, 1, 1, &page);
1702 base = kmap_atomic(page);
1704 kunmap_atomic(base);
1705 set_page_dirty_lock(page);
1710 static int log_write(void __user *log_base,
1711 u64 write_address, u64 write_length)
1713 u64 write_page = write_address / VHOST_PAGE_SIZE;
1718 write_length += write_address % VHOST_PAGE_SIZE;
1720 u64 base = (u64)(unsigned long)log_base;
1721 u64 log = base + write_page / 8;
1722 int bit = write_page % 8;
1723 if ((u64)(unsigned long)log != log)
1725 r = set_bit_to_user(bit, (void __user *)(unsigned long)log);
1728 if (write_length <= VHOST_PAGE_SIZE)
1730 write_length -= VHOST_PAGE_SIZE;
1736 int vhost_log_write(struct vhost_virtqueue *vq, struct vhost_log *log,
1737 unsigned int log_num, u64 len)
1741 /* Make sure data written is seen before log. */
1743 for (i = 0; i < log_num; ++i) {
1744 u64 l = min(log[i].len, len);
1745 r = log_write(vq->log_base, log[i].addr, l);
1751 eventfd_signal(vq->log_ctx, 1);
1755 /* Length written exceeds what we have stored. This is a bug. */
1759 EXPORT_SYMBOL_GPL(vhost_log_write);
1761 static int vhost_update_used_flags(struct vhost_virtqueue *vq)
1764 if (vhost_put_user(vq, cpu_to_vhost16(vq, vq->used_flags),
1765 &vq->used->flags) < 0)
1767 if (unlikely(vq->log_used)) {
1768 /* Make sure the flag is seen before log. */
1770 /* Log used flag write. */
1771 used = &vq->used->flags;
1772 log_write(vq->log_base, vq->log_addr +
1773 (used - (void __user *)vq->used),
1774 sizeof vq->used->flags);
1776 eventfd_signal(vq->log_ctx, 1);
1781 static int vhost_update_avail_event(struct vhost_virtqueue *vq, u16 avail_event)
1783 if (vhost_put_user(vq, cpu_to_vhost16(vq, vq->avail_idx),
1784 vhost_avail_event(vq)))
1786 if (unlikely(vq->log_used)) {
1788 /* Make sure the event is seen before log. */
1790 /* Log avail event write */
1791 used = vhost_avail_event(vq);
1792 log_write(vq->log_base, vq->log_addr +
1793 (used - (void __user *)vq->used),
1794 sizeof *vhost_avail_event(vq));
1796 eventfd_signal(vq->log_ctx, 1);
1801 int vhost_vq_init_access(struct vhost_virtqueue *vq)
1803 __virtio16 last_used_idx;
1805 bool is_le = vq->is_le;
1807 if (!vq->private_data)
1810 vhost_init_is_le(vq);
1812 r = vhost_update_used_flags(vq);
1815 vq->signalled_used_valid = false;
1817 !access_ok(VERIFY_READ, &vq->used->idx, sizeof vq->used->idx)) {
1821 r = vhost_get_used(vq, last_used_idx, &vq->used->idx);
1823 vq_err(vq, "Can't access used idx at %p\n",
1827 vq->last_used_idx = vhost16_to_cpu(vq, last_used_idx);
1834 EXPORT_SYMBOL_GPL(vhost_vq_init_access);
1836 static int translate_desc(struct vhost_virtqueue *vq, u64 addr, u32 len,
1837 struct iovec iov[], int iov_size, int access)
1839 const struct vhost_umem_node *node;
1840 struct vhost_dev *dev = vq->dev;
1841 struct vhost_umem *umem = dev->iotlb ? dev->iotlb : dev->umem;
1846 while ((u64)len > s) {
1848 if (unlikely(ret >= iov_size)) {
1853 node = vhost_umem_interval_tree_iter_first(&umem->umem_tree,
1854 addr, addr + len - 1);
1855 if (node == NULL || node->start > addr) {
1856 if (umem != dev->iotlb) {
1862 } else if (!(node->perm & access)) {
1868 size = node->size - addr + node->start;
1869 _iov->iov_len = min((u64)len - s, size);
1870 _iov->iov_base = (void __user *)(unsigned long)
1871 (node->userspace_addr + addr - node->start);
1878 vhost_iotlb_miss(vq, addr, access);
1882 /* Each buffer in the virtqueues is actually a chain of descriptors. This
1883 * function returns the next descriptor in the chain,
1884 * or -1U if we're at the end. */
1885 static unsigned next_desc(struct vhost_virtqueue *vq, struct vring_desc *desc)
1889 /* If this descriptor says it doesn't chain, we're done. */
1890 if (!(desc->flags & cpu_to_vhost16(vq, VRING_DESC_F_NEXT)))
1893 /* Check they're not leading us off end of descriptors. */
1894 next = vhost16_to_cpu(vq, READ_ONCE(desc->next));
1898 static int get_indirect(struct vhost_virtqueue *vq,
1899 struct iovec iov[], unsigned int iov_size,
1900 unsigned int *out_num, unsigned int *in_num,
1901 struct vhost_log *log, unsigned int *log_num,
1902 struct vring_desc *indirect)
1904 struct vring_desc desc;
1905 unsigned int i = 0, count, found = 0;
1906 u32 len = vhost32_to_cpu(vq, indirect->len);
1907 struct iov_iter from;
1911 if (unlikely(len % sizeof desc)) {
1912 vq_err(vq, "Invalid length in indirect descriptor: "
1913 "len 0x%llx not multiple of 0x%zx\n",
1914 (unsigned long long)len,
1919 ret = translate_desc(vq, vhost64_to_cpu(vq, indirect->addr), len, vq->indirect,
1920 UIO_MAXIOV, VHOST_ACCESS_RO);
1921 if (unlikely(ret < 0)) {
1923 vq_err(vq, "Translation failure %d in indirect.\n", ret);
1926 iov_iter_init(&from, READ, vq->indirect, ret, len);
1928 /* We will use the result as an address to read from, so most
1929 * architectures only need a compiler barrier here. */
1930 read_barrier_depends();
1932 count = len / sizeof desc;
1933 /* Buffers are chained via a 16 bit next field, so
1934 * we can have at most 2^16 of these. */
1935 if (unlikely(count > USHRT_MAX + 1)) {
1936 vq_err(vq, "Indirect buffer length too big: %d\n",
1942 unsigned iov_count = *in_num + *out_num;
1943 if (unlikely(++found > count)) {
1944 vq_err(vq, "Loop detected: last one at %u "
1945 "indirect size %u\n",
1949 if (unlikely(!copy_from_iter_full(&desc, sizeof(desc), &from))) {
1950 vq_err(vq, "Failed indirect descriptor: idx %d, %zx\n",
1951 i, (size_t)vhost64_to_cpu(vq, indirect->addr) + i * sizeof desc);
1954 if (unlikely(desc.flags & cpu_to_vhost16(vq, VRING_DESC_F_INDIRECT))) {
1955 vq_err(vq, "Nested indirect descriptor: idx %d, %zx\n",
1956 i, (size_t)vhost64_to_cpu(vq, indirect->addr) + i * sizeof desc);
1960 if (desc.flags & cpu_to_vhost16(vq, VRING_DESC_F_WRITE))
1961 access = VHOST_ACCESS_WO;
1963 access = VHOST_ACCESS_RO;
1965 ret = translate_desc(vq, vhost64_to_cpu(vq, desc.addr),
1966 vhost32_to_cpu(vq, desc.len), iov + iov_count,
1967 iov_size - iov_count, access);
1968 if (unlikely(ret < 0)) {
1970 vq_err(vq, "Translation failure %d indirect idx %d\n",
1974 /* If this is an input descriptor, increment that count. */
1975 if (access == VHOST_ACCESS_WO) {
1977 if (unlikely(log)) {
1978 log[*log_num].addr = vhost64_to_cpu(vq, desc.addr);
1979 log[*log_num].len = vhost32_to_cpu(vq, desc.len);
1983 /* If it's an output descriptor, they're all supposed
1984 * to come before any input descriptors. */
1985 if (unlikely(*in_num)) {
1986 vq_err(vq, "Indirect descriptor "
1987 "has out after in: idx %d\n", i);
1992 } while ((i = next_desc(vq, &desc)) != -1);
1996 /* This looks in the virtqueue and for the first available buffer, and converts
1997 * it to an iovec for convenient access. Since descriptors consist of some
1998 * number of output then some number of input descriptors, it's actually two
1999 * iovecs, but we pack them into one and note how many of each there were.
2001 * This function returns the descriptor number found, or vq->num (which is
2002 * never a valid descriptor number) if none was found. A negative code is
2003 * returned on error. */
2004 int vhost_get_vq_desc(struct vhost_virtqueue *vq,
2005 struct iovec iov[], unsigned int iov_size,
2006 unsigned int *out_num, unsigned int *in_num,
2007 struct vhost_log *log, unsigned int *log_num)
2009 struct vring_desc desc;
2010 unsigned int i, head, found = 0;
2012 __virtio16 avail_idx;
2013 __virtio16 ring_head;
2016 /* Check it isn't doing very strange things with descriptor numbers. */
2017 last_avail_idx = vq->last_avail_idx;
2019 if (vq->avail_idx == vq->last_avail_idx) {
2020 if (unlikely(vhost_get_avail(vq, avail_idx, &vq->avail->idx))) {
2021 vq_err(vq, "Failed to access avail idx at %p\n",
2025 vq->avail_idx = vhost16_to_cpu(vq, avail_idx);
2027 if (unlikely((u16)(vq->avail_idx - last_avail_idx) > vq->num)) {
2028 vq_err(vq, "Guest moved used index from %u to %u",
2029 last_avail_idx, vq->avail_idx);
2033 /* If there's nothing new since last we looked, return
2036 if (vq->avail_idx == last_avail_idx)
2039 /* Only get avail ring entries after they have been
2045 /* Grab the next descriptor number they're advertising, and increment
2046 * the index we've seen. */
2047 if (unlikely(vhost_get_avail(vq, ring_head,
2048 &vq->avail->ring[last_avail_idx & (vq->num - 1)]))) {
2049 vq_err(vq, "Failed to read head: idx %d address %p\n",
2051 &vq->avail->ring[last_avail_idx % vq->num]);
2055 head = vhost16_to_cpu(vq, ring_head);
2057 /* If their number is silly, that's an error. */
2058 if (unlikely(head >= vq->num)) {
2059 vq_err(vq, "Guest says index %u > %u is available",
2064 /* When we start there are none of either input nor output. */
2065 *out_num = *in_num = 0;
2071 unsigned iov_count = *in_num + *out_num;
2072 if (unlikely(i >= vq->num)) {
2073 vq_err(vq, "Desc index is %u > %u, head = %u",
2077 if (unlikely(++found > vq->num)) {
2078 vq_err(vq, "Loop detected: last one at %u "
2079 "vq size %u head %u\n",
2083 ret = vhost_copy_from_user(vq, &desc, vq->desc + i,
2085 if (unlikely(ret)) {
2086 vq_err(vq, "Failed to get descriptor: idx %d addr %p\n",
2090 if (desc.flags & cpu_to_vhost16(vq, VRING_DESC_F_INDIRECT)) {
2091 ret = get_indirect(vq, iov, iov_size,
2093 log, log_num, &desc);
2094 if (unlikely(ret < 0)) {
2096 vq_err(vq, "Failure detected "
2097 "in indirect descriptor at idx %d\n", i);
2103 if (desc.flags & cpu_to_vhost16(vq, VRING_DESC_F_WRITE))
2104 access = VHOST_ACCESS_WO;
2106 access = VHOST_ACCESS_RO;
2107 ret = translate_desc(vq, vhost64_to_cpu(vq, desc.addr),
2108 vhost32_to_cpu(vq, desc.len), iov + iov_count,
2109 iov_size - iov_count, access);
2110 if (unlikely(ret < 0)) {
2112 vq_err(vq, "Translation failure %d descriptor idx %d\n",
2116 if (access == VHOST_ACCESS_WO) {
2117 /* If this is an input descriptor,
2118 * increment that count. */
2120 if (unlikely(log)) {
2121 log[*log_num].addr = vhost64_to_cpu(vq, desc.addr);
2122 log[*log_num].len = vhost32_to_cpu(vq, desc.len);
2126 /* If it's an output descriptor, they're all supposed
2127 * to come before any input descriptors. */
2128 if (unlikely(*in_num)) {
2129 vq_err(vq, "Descriptor has out after in: "
2135 } while ((i = next_desc(vq, &desc)) != -1);
2137 /* On success, increment avail index. */
2138 vq->last_avail_idx++;
2140 /* Assume notifications from guest are disabled at this point,
2141 * if they aren't we would need to update avail_event index. */
2142 BUG_ON(!(vq->used_flags & VRING_USED_F_NO_NOTIFY));
2145 EXPORT_SYMBOL_GPL(vhost_get_vq_desc);
2147 /* Reverse the effect of vhost_get_vq_desc. Useful for error handling. */
2148 void vhost_discard_vq_desc(struct vhost_virtqueue *vq, int n)
2150 vq->last_avail_idx -= n;
2152 EXPORT_SYMBOL_GPL(vhost_discard_vq_desc);
2154 /* After we've used one of their buffers, we tell them about it. We'll then
2155 * want to notify the guest, using eventfd. */
2156 int vhost_add_used(struct vhost_virtqueue *vq, unsigned int head, int len)
2158 struct vring_used_elem heads = {
2159 cpu_to_vhost32(vq, head),
2160 cpu_to_vhost32(vq, len)
2163 return vhost_add_used_n(vq, &heads, 1);
2165 EXPORT_SYMBOL_GPL(vhost_add_used);
2167 static int __vhost_add_used_n(struct vhost_virtqueue *vq,
2168 struct vring_used_elem *heads,
2171 struct vring_used_elem __user *used;
2175 start = vq->last_used_idx & (vq->num - 1);
2176 used = vq->used->ring + start;
2178 if (vhost_put_user(vq, heads[0].id, &used->id)) {
2179 vq_err(vq, "Failed to write used id");
2182 if (vhost_put_user(vq, heads[0].len, &used->len)) {
2183 vq_err(vq, "Failed to write used len");
2186 } else if (vhost_copy_to_user(vq, used, heads, count * sizeof *used)) {
2187 vq_err(vq, "Failed to write used");
2190 if (unlikely(vq->log_used)) {
2191 /* Make sure data is seen before log. */
2193 /* Log used ring entry write. */
2194 log_write(vq->log_base,
2196 ((void __user *)used - (void __user *)vq->used),
2197 count * sizeof *used);
2199 old = vq->last_used_idx;
2200 new = (vq->last_used_idx += count);
2201 /* If the driver never bothers to signal in a very long while,
2202 * used index might wrap around. If that happens, invalidate
2203 * signalled_used index we stored. TODO: make sure driver
2204 * signals at least once in 2^16 and remove this. */
2205 if (unlikely((u16)(new - vq->signalled_used) < (u16)(new - old)))
2206 vq->signalled_used_valid = false;
2210 /* After we've used one of their buffers, we tell them about it. We'll then
2211 * want to notify the guest, using eventfd. */
2212 int vhost_add_used_n(struct vhost_virtqueue *vq, struct vring_used_elem *heads,
2217 start = vq->last_used_idx & (vq->num - 1);
2218 n = vq->num - start;
2220 r = __vhost_add_used_n(vq, heads, n);
2226 r = __vhost_add_used_n(vq, heads, count);
2228 /* Make sure buffer is written before we update index. */
2230 if (vhost_put_user(vq, cpu_to_vhost16(vq, vq->last_used_idx),
2232 vq_err(vq, "Failed to increment used idx");
2235 if (unlikely(vq->log_used)) {
2236 /* Make sure used idx is seen before log. */
2238 /* Log used index update. */
2239 log_write(vq->log_base,
2240 vq->log_addr + offsetof(struct vring_used, idx),
2241 sizeof vq->used->idx);
2243 eventfd_signal(vq->log_ctx, 1);
2247 EXPORT_SYMBOL_GPL(vhost_add_used_n);
2249 static bool vhost_notify(struct vhost_dev *dev, struct vhost_virtqueue *vq)
2254 /* Flush out used index updates. This is paired
2255 * with the barrier that the Guest executes when enabling
2259 if (vhost_has_feature(vq, VIRTIO_F_NOTIFY_ON_EMPTY) &&
2260 unlikely(vq->avail_idx == vq->last_avail_idx))
2263 if (!vhost_has_feature(vq, VIRTIO_RING_F_EVENT_IDX)) {
2265 if (vhost_get_avail(vq, flags, &vq->avail->flags)) {
2266 vq_err(vq, "Failed to get flags");
2269 return !(flags & cpu_to_vhost16(vq, VRING_AVAIL_F_NO_INTERRUPT));
2271 old = vq->signalled_used;
2272 v = vq->signalled_used_valid;
2273 new = vq->signalled_used = vq->last_used_idx;
2274 vq->signalled_used_valid = true;
2279 if (vhost_get_avail(vq, event, vhost_used_event(vq))) {
2280 vq_err(vq, "Failed to get used event idx");
2283 return vring_need_event(vhost16_to_cpu(vq, event), new, old);
2286 /* This actually signals the guest, using eventfd. */
2287 void vhost_signal(struct vhost_dev *dev, struct vhost_virtqueue *vq)
2289 /* Signal the Guest tell them we used something up. */
2290 if (vq->call_ctx && vhost_notify(dev, vq))
2291 eventfd_signal(vq->call_ctx, 1);
2293 EXPORT_SYMBOL_GPL(vhost_signal);
2295 /* And here's the combo meal deal. Supersize me! */
2296 void vhost_add_used_and_signal(struct vhost_dev *dev,
2297 struct vhost_virtqueue *vq,
2298 unsigned int head, int len)
2300 vhost_add_used(vq, head, len);
2301 vhost_signal(dev, vq);
2303 EXPORT_SYMBOL_GPL(vhost_add_used_and_signal);
2305 /* multi-buffer version of vhost_add_used_and_signal */
2306 void vhost_add_used_and_signal_n(struct vhost_dev *dev,
2307 struct vhost_virtqueue *vq,
2308 struct vring_used_elem *heads, unsigned count)
2310 vhost_add_used_n(vq, heads, count);
2311 vhost_signal(dev, vq);
2313 EXPORT_SYMBOL_GPL(vhost_add_used_and_signal_n);
2315 /* return true if we're sure that avaiable ring is empty */
2316 bool vhost_vq_avail_empty(struct vhost_dev *dev, struct vhost_virtqueue *vq)
2318 __virtio16 avail_idx;
2321 if (vq->avail_idx != vq->last_avail_idx)
2324 r = vhost_get_avail(vq, avail_idx, &vq->avail->idx);
2327 vq->avail_idx = vhost16_to_cpu(vq, avail_idx);
2329 return vq->avail_idx == vq->last_avail_idx;
2331 EXPORT_SYMBOL_GPL(vhost_vq_avail_empty);
2333 /* OK, now we need to know about added descriptors. */
2334 bool vhost_enable_notify(struct vhost_dev *dev, struct vhost_virtqueue *vq)
2336 __virtio16 avail_idx;
2339 if (!(vq->used_flags & VRING_USED_F_NO_NOTIFY))
2341 vq->used_flags &= ~VRING_USED_F_NO_NOTIFY;
2342 if (!vhost_has_feature(vq, VIRTIO_RING_F_EVENT_IDX)) {
2343 r = vhost_update_used_flags(vq);
2345 vq_err(vq, "Failed to enable notification at %p: %d\n",
2346 &vq->used->flags, r);
2350 r = vhost_update_avail_event(vq, vq->avail_idx);
2352 vq_err(vq, "Failed to update avail event index at %p: %d\n",
2353 vhost_avail_event(vq), r);
2357 /* They could have slipped one in as we were doing that: make
2358 * sure it's written, then check again. */
2360 r = vhost_get_avail(vq, avail_idx, &vq->avail->idx);
2362 vq_err(vq, "Failed to check avail idx at %p: %d\n",
2363 &vq->avail->idx, r);
2367 return vhost16_to_cpu(vq, avail_idx) != vq->avail_idx;
2369 EXPORT_SYMBOL_GPL(vhost_enable_notify);
2371 /* We don't need to be notified again. */
2372 void vhost_disable_notify(struct vhost_dev *dev, struct vhost_virtqueue *vq)
2376 if (vq->used_flags & VRING_USED_F_NO_NOTIFY)
2378 vq->used_flags |= VRING_USED_F_NO_NOTIFY;
2379 if (!vhost_has_feature(vq, VIRTIO_RING_F_EVENT_IDX)) {
2380 r = vhost_update_used_flags(vq);
2382 vq_err(vq, "Failed to enable notification at %p: %d\n",
2383 &vq->used->flags, r);
2386 EXPORT_SYMBOL_GPL(vhost_disable_notify);
2388 /* Create a new message. */
2389 struct vhost_msg_node *vhost_new_msg(struct vhost_virtqueue *vq, int type)
2391 struct vhost_msg_node *node = kmalloc(sizeof *node, GFP_KERNEL);
2395 /* Make sure all padding within the structure is initialized. */
2396 memset(&node->msg, 0, sizeof node->msg);
2398 node->msg.type = type;
2401 EXPORT_SYMBOL_GPL(vhost_new_msg);
2403 void vhost_enqueue_msg(struct vhost_dev *dev, struct list_head *head,
2404 struct vhost_msg_node *node)
2406 spin_lock(&dev->iotlb_lock);
2407 list_add_tail(&node->node, head);
2408 spin_unlock(&dev->iotlb_lock);
2410 wake_up_interruptible_poll(&dev->wait, EPOLLIN | EPOLLRDNORM);
2412 EXPORT_SYMBOL_GPL(vhost_enqueue_msg);
2414 struct vhost_msg_node *vhost_dequeue_msg(struct vhost_dev *dev,
2415 struct list_head *head)
2417 struct vhost_msg_node *node = NULL;
2419 spin_lock(&dev->iotlb_lock);
2420 if (!list_empty(head)) {
2421 node = list_first_entry(head, struct vhost_msg_node,
2423 list_del(&node->node);
2425 spin_unlock(&dev->iotlb_lock);
2429 EXPORT_SYMBOL_GPL(vhost_dequeue_msg);
2432 static int __init vhost_init(void)
2437 static void __exit vhost_exit(void)
2441 module_init(vhost_init);
2442 module_exit(vhost_exit);
2444 MODULE_VERSION("0.0.1");
2445 MODULE_LICENSE("GPL v2");
2446 MODULE_AUTHOR("Michael S. Tsirkin");
2447 MODULE_DESCRIPTION("Host kernel accelerator for virtio");